This is the module you use rather than read. Every template here is copy-paste ready, with inline guidance explaining what each section is for and a filled example showing one plausible way to complete it. The examples are illustrative: replace them with your own product, your own numbers, and your own names before anyone else sees the document. One honest warning: a template is only worth using if you delete the parts that do not apply to you. A half-filled document with six empty headings does more damage than a short one that answers the questions your team actually faces.
- After this module you can write a PRD for an AI feature that states a quality bar, a fallback behaviour, and a cost ceiling, so engineering and product argue about the right things early.
- After this module you can produce an eval plan that names the dataset, the metric, the threshold, and the person who owns the number.
- After this module you can record a model choice as a decision record, including the trigger that would make you reverse it.
- After this module you can run an AI launch review and an AI incident postmortem without falling back on "the model was wrong" as an explanation.
- After this module you can send a weekly status update that a director can read in two minutes and act on.
| # | Page | Time | What you get |
|---|---|---|---|
| 1 | AI feature PRD | 15 minutes | The quality bar, the fallback, and the cost ceiling that a normal PRD leaves out. |
| 2 | Eval plan | 15 minutes | One page that turns "is it good?" into a number and an owner. |
| 3 | Model selection decision record | 10 minutes | An ADR for a model choice, with the reversal trigger that matters most. |
| 4 | AI launch checklist | 15 minutes | The pre-launch questions your existing checklist does not ask. |
| 5 | AI incident postmortem | 15 minutes | Why the model is never the root cause, and the eval case that must come out of it. |
| 6 | Weekly AI status update | 10 minutes | Eval movement, cost trend, top risks, decisions needed. |
About 1 hour 20 minutes to read, and they are meant to be copied rather than read.
Copy the template into wherever your team already keeps documents, then cut it down. Delete every heading you cannot fill in honestly today, and note the ones you deleted on purpose so a reviewer knows the gap is deliberate rather than missed. If you find yourself fighting the structure, the structure is wrong for your situation, so change it. These are starting points that save you a blank page, not a process to comply with.
Last reviewed: July 2026
Next: Repository home
Related: Managing AI teams | Evaluation fundamentals | Responsible AI in practice